State and Parameter Estimation Approach for Online Adaptation of Data-Driven Models

Data driven models, such as neural networks. are widely used to model dynamical systems. They are relatively easy to develop, and are able to represent dynamical systems using algebraic relationships. However, it is not easy to update these models as plant behaviour changes over a period of time. Complete retraining is possible, but it is too time consuming and needs huge amount of data which may not be available during online operation. The current project will investigate state and parameter estimation techniques from state estimation literature to update parameters of a data driven model in a recursive, and online manner. Towards this end, sensitivity analysis is used to map parameters of a first principles model to data driven model. Principal component analysis (PCA) is then used to reduce the dimensionality of the parameters of the data driven model. In online phase, classical random walk model based parameter update procedure is then used within state estimation framework to update the parameters of the data driven model along with the states of the process.

The work is continuation of previous MTP works.

Name of Faculty